Remittances and the Brain Drain Revisited: The Microdata Show That More Educated Migrants Remit More

Remittances and the Brain Drain Revisited: The Microdata Show That More Educated Migrants Remit More Albert Bollard, David McKenzie, Melanie Morten, and Hillel Rapoport Two of the most salient trends in migration and development over the last two decades are the large rise in remittances and in the flow of skilled migrants. However, recent literature based on cross-country regressions has claimed that more educated migrants remit less, leading to concerns that further increases in skilled migration will impede remittance growth. Microdata from surveys of immigrants in 11 major destination countries are used to revisit the relationship between education and remitting behavior. The data show a mixed pattern between education and the likelihood of remitting, and a strong positive relationship between education and amount remitted (intensive margin), conditional on remitting at all (extensive margin). Combining these intensive and extensive margins yields an overall positive effect of education on the amount remitted for the pooled sample, with heterogeneous results across destinations. The microdata allow investigation of why the more educated remit more, showing that the higher income earned by migrants, rather than family characteristics, explains much of the higher remittances. remittances, migration, brain drain, education JEL codes: O15, F22, J61 Two of the most salient trends in migration and development over the last two decades are the large rise in remittances and in the flow of skilled migrants. Officially recorded remittances to developing countries have more than tripled over the last decade, rising from $85 billion in 2000 to $305 billion in 2008 Albert Bollard (abollard@stanford.edu) is a PhD student at Stanford University. David McKenzie (dmckenzie@worldbank.org; corresponding author) is a senior economist in the Finance and Private Sector Research Unit of the Development Research Group at the World Bank. Melanie Morten (morten@yale.edu) is a PhD student at Yale University. Hillel Rapoport (hillel@mail.biu.ac.il) is professor of economics at Bar Ilan University and at EQUIPPE, University of Lille and is currently a visiting research fellow at the Center for International Development at Harvard University. The authors are grateful for funding for this project from the Agence Française de Développement (AFD). They thank all the individuals and organizations that graciously shared their surveys of immigrants, and they are grateful to Michael Clemens, participants at the 2 nd Migration and Development Conference held at the World Bank in September 2009, three anonymous referees, and the journal editor for helpful comments. All opinions are those of the authors and do not necessarily represent those of AFD or the World Bank. A supplemental appendix to this article is available at http://wber.oxfordjournals.org. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 1, pp. 132 156 doi:10.1093/wber/lhr013 Advance Access Publication May 12, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 132

Bollard, McKenzie, Morten, and Rapoport 133 (World Bank 2008, 2009). The number of highly educated migrants from developing countries residing in Organisation for Economic Co-operation and Development (OECD) countries doubled over 1990 2000 (Docquier and Marfouk 2005) and likely has grown since then as developed countries have increasingly pursued skill-selective immigration policies. 1 However, despite this positive association at the global level between rising remittances and rising high-skill migration, there are concerns stemming from the belief that more educated individuals may remit less that increasingly skill-selective immigration policies may slow or even end the rise in remittances. This belief is taken as fact by many; for example, an OECD (2007, p. 11) report says that low skilled migrants tend to send more money home. The main empirical evidence to support this assertion across a range of countries comes from two recent studies (Faini 2007; Niimi, Özden, and Schiff forthcoming) whose cross-country macroeconomic analyses find that the highly skilled (defined as those with tertiary education) remit less. Yet there are many reasons to question the results of these cross-country estimations. Both studies relate the amount of remittances received at a country level to the share of migrants with tertiary education, at best telling us whether countries that send a larger share of highly skilled migrants receive less or more in remittances than countries that send fewer skilled migrants. The studies do not answer the factual question of whether more educated migrants remit more or less. There are a host of differences across countries that could cause a spurious relationship to appear between remittances and skill level across countries. For example, if poverty is a constraint to both migration and education, richer developing countries might be able to send more migrants (yielding more remittances) and those migrants might also have more schooling. Faini (2007) treats the share of migrants who are skilled as exogenous. Niimi, Özden, and Schiff (forthcoming) try to instrument for the education mix of migrants, but their instruments seem unlikely to satisfy the exclusion restrictions. For example, public spending on education is likely a function of a country s overall institutional and economic development, which should independently affect the incentive to remit; migrants might send money to overcome poor public spending or for investment when complementary infrastructure and institutions are in place. This article revisits the relationship between remittances and education level using microdata that permit computing the association between a migrant s education level and remitting behavior. The authors assembled the most comprehensive micro-level database on remitting behavior currently available, comprising data on 33,000 immigrants from developing countries from 14 surveys 1. In contrast, the number of low-skill migrants ( primary education or less) increased only 15 percent over the period. Immigration to OECD countries (as defined by the number of foreign born) was estimated at 90 million in 2000, about half of total world migration. Of the 90 million immigrants, 60 million were ages 25 or older and were split equally across education categories (primary, secondary, and tertiary; Docquier and Marfouk 2005).

134 THE WORLD BANK ECONOMIC REVIEW in 11 OECD destination countries. The analysis begins by establishing the factual relationship between the propensity to remit and education. No attempt is made to estimate the causal impact of education on remittances. 2 From a policy perspective, the concern is whether migration policies that shift the education composition of migrants affect remittances, not whether education policies that change how much education individuals have affects remittances. Microdata enables asking whether more educated individuals are more or less likely to remit (the extensive margin) and whether they send more or less remittances if they do remit (the intensive margin). A mixed association is found between education and remittances at the extensive margin, and a strong positive relationship at the intensive margin. Combining both the extensive and intensive margins reveals that, at least in this large sample, more educated migrants do remit significantly more migrants with a university degree remit $300 more yearly than migrants without a university degree, where the mean annual remittance over the entire sample is $730. The article is organized as follows. Section I summarizes several theories of remitting behavior and the predictions they give for the relationship between education and remittances. Section II then describes the dataset of immigrant surveys with remittances. Section III provides results, and section IV draws some implications. I. THEORETICAL B ACKGROUND Theoretically, there are several reasons to believe that there will be differences in the remitting patterns of highly skilled and less-skilled emigrants. However, a priori, it is not clear which direction will dominate and thus whether the highly skilled will remit more or less on average. On the one hand, several factors tend to lead highly skilled migrants to be more likely to remit and to send a larger amount of remittances. First, highly skilled individuals are likely to earn more as migrants, potentially increasing the amount they can remit. Second, their education may have been funded by family members in the home country, with remittances serving as repayment. Third, skilled migrants are less likely to be illegal migrants and more likely to have bank accounts, lowering the financial transaction costs of remitting. On the other hand, several other factors might lead highly skilled migrants to be less likely to remit and to remit less. First, highly skilled migrants may be more likely to migrate with their entire household, so they would not have to send remittances in order to share their earnings with their household. Second, they might come from richer households, which have less need for remittances to alleviate liquidity constraints. Third, they might have less intention of returning to their home country, reducing the role of remittances as a way of maintaining prestige and ties to the home community. 2. Convincing instruments are lacking to identify this impact.

Bollard, McKenzie, Morten, and Rapoport 135 Before turning to the empirical analysis, it is useful to clarify the theoretical relationship between education and remittances and the implied testable predictions about education. This will allow identifying the role of several variables that, once interacted with education and various possible motivations to remit, have the potential to explain differences in remitting behavior by education level. The discussion is limited to three possible motives for remittances: altruism, exchange, and investment. These were selected for general empirical relevance and as the motives through which education is most likely to affect remittances. 3 Altruism Altruistic preferences are generally captured by weighting one s own (the migrant s) and others (relatives) consumption in an individual utility function, with weights reflecting the individual s degree of altruism, which can itself depend on the closeness among the relatives considered (both family and geographic proximity). For given weights and initial distribution of income, altruistic individuals maximize their utility by transferring (remitting) income so as to reach the desired distribution between themselves and the beneficiaries of their altruism. Altruistic transfers take place if pretransfer income differences are sufficiently large or altruism is strong enough and increases with the donor s income (the extensive margin) and decreases with the recipients income (the intensive margin) What does this basic theoretical framework imply for the comparative remitting behavior of highly educated and less well educated migrants? First, educated migrants tend to earn more, which all else equal should induce more remittances (at both margins). Second, the conventional wisdom is that educated migrants tend to have more family members with them because of a higher propensity to move with their immediate family, which all else equal should lower remittances. 4 Methodologically, this suggests that the location and composition of the family (which fraction of the family accompanies the migrant and which fraction stays in the country of origin) is jointly determined with remittances. This makes it difficult to estimate the causal impact of family composition on remittances. Instead, the analysis simply looks at whether differences in remitting patterns by education level disappear when they are conditioned on family composition. Empirically, the analysis will show that while less educated migrants have more relatives in the home country, they also have larger households and more relatives with them in the destination country. 3. See Rapoport and Docquier (2006) for a comprehensive survey of the economic literature on migrants remittances. 4. In this basic framework, education has no impact beyond its effect on the migrants income and family size, composition, and location, and altruistic preferences are independent of education.

136 THE WORLD BANK ECONOMIC REVIEW Exchange and Investment Motives There are many situations of Pareto-improving exchanges in which remittances buy various types of services, such as taking care of the migrant s assets (land and cattle, for example) or relatives (children, elderly parents) at home. Such motivations are generally a sign of temporary migration and signal a migrant s intention to return. In such exchanges, there is a participation constraint determined by each partner s external options, with the exact division of the pie (or surplus) to be shared depending on each partner s bargaining power. How does education interact with such exchange motives? Two directions emerge from the short discussion above: one through the effect of education on intentions to return, and another through education s effect on threat points and bargaining powers. The conventional wisdom is that migrants with higher education have less intention (and propensity) to return than do migrants with lower education (see Faini 2007), because they are better integrated or can obtain permanent resident status more easily. If that is the case, more educated migrants should transfer less for an exchange motive, reflecting their lower propensity to return. 5 What about bargaining powers? Exchange models allow for different possible contractual arrangements reflecting the parties outside options and bargaining powers (see, for example, Cox 1987; Cox, Eser, and Jimenez 1998). This has two complementary implications for education as a determinant of remittances in an exchange model. First, to the extent that education is associated with higher income, this relationship is likely to increase a migrant s willingness to pay, leading to higher remittances; and second, to the extent that educated migrants come from more affluent families, this relationship is likely to increase the receiving household s bargaining power, also leading to higher remittances. On the whole, an exchange motive therefore predicts that education will have an ambiguous effect on remittances, with the sign of the effect depending on whether return intentions or bargaining issues matter more to remittance behavior. The investment motive can be seen as a particular exchange of services in a context of imperfect credit markets. In such a context, remittances can be seen as part of an implicit migration contract between migrant and family, allowing the family access to higher income (investment motive) or less volatile income (insurance motive; Stark 1991). Since the insurance motive does not in theory give rise to clear differences in transfer behavior between highly educated and less educated migrants, the focus here is on the investment motive. The amount of investment financed by the family may include the physical costs (such as transportation) and informational costs of migration, as well as education expenditures, and repayment of this implicit loan through remittances is obviously expected to depend on the magnitude of the loan. Thus, the 5. Again, as shown later in the article, this conventional wisdom is not supported by the data; exchange motives are equally relevant for highly educated and less educated migrants as far as return intentions are concerned.

Bollard, McKenzie, Morten, and Rapoport 137 investment motive clearly predicts that, all else equal, more educated migrants should remit more to compensate the family for the additional education expenditures incurred. Summary of Predictions Both the altruistic and the exchange motives for remittances yield unclear theoretical predictions as to whether more educated migrants remit more or less than do less education migrants. Once migrants incomes are controlled for, their education level should not play a role under the altruistic hypothesis (assuming preferences are exogenous to education) except for its effect on the spatial distribution of the family. As already noted, the conventional wisdom here is that the highly educated tend to move with their immediate family, which would lower remittances. Similarly, education is expected to lower remittances under the exchange hypothesis if educated migrants have lower propensities to return; bargaining mechanisms work in the other direction and should translate into higher remittances, with the sign of the total expected effect being theoretically uncertain. Finally, education is likely to have a clear positive impact on remittances under the investment hypothesis. Given these expected mechanisms and the fact that the descriptive statistics for the sample do not support the conjecture that more educated migrants have a substantially higher propensity to move with their family or a substantially lower propensity to return, the other forces at work should be expected to dominate, so that migrants with more education would remit more, which is indeed what the analysis shows. II. DATA The micro-level database on remitting behavior created for this study is the most comprehensive available, comprising data on 33,000 immigrants from developing countries derived from 14 surveys in 11 OECD destination countries that were the destination for 79 percent of global migrants to OECD countries in 2000 (Docquier and Marfouk 2005). The focus on destination country data sources enables looking directly at the relationship between education and remittance sending behavior by analyzing the migrants decision to remit. It also permits capturing the remittance behavior of individuals who emigrate with their entire household; using household surveys from the remittance receiving countries would typically miss such individuals. Since more educated individuals are believed to be more likely to emigrate with their entire household than less educated individuals (Faini 2007), using surveys from migrant sending countries would not be appropriate for examining the relationship between remittances and education. Most of the empirical literature on immigrants uses data from censuses or labor force surveys, but neither contains information on remittances. That

138 THE WORLD BANK ECONOMIC REVIEW requires special purpose surveys of immigrants. The authors pulled together all publicly available datasets they were aware of 6 and six additional surveys that are not publicly available but that other researchers generously shared. Table 1 provides an overview of the database of migrants, summarizing the datasets, sample population, and survey methodology. Full details of the source of each dataset are in the supplemental appendix, available online at http://wber. oxfordjournals.org/. The database covers a wide range of populations. It includes both nationally representative surveys, such as the New Immigrant Survey (NIS) in the United States (drawn from green card recipients) and the Spanish National Survey of Immigrants (ENI), which draws on a neighborhood sampling frame, as well as surveys focusing on specific migrant communities within the recipient country, such as the Black/Minority Ethnic Survey (BME) in the United Kingdom and the Belgium International Remittance Senders Household Survey (IRSHS) of immigrants from the Democratic Republic of Congo, Nigeria, and Senegal. In all cases, the database includes only migrants who were born in developing countries. 7 For each country dataset, comparable covariates were constructed to measure household income, remittance behavior, family composition, and demographic characteristics. Remittances are typically measured at the household rather than individual level. The level of analysis is therefore the household, and variables are defined at this level whenever possible for example, by taking the highest level of schooling achieved by any adult migrant in the household. All financial values are reported in 2003 U.S. dollars. In addition, any reported annual remittances that are more than twice annual household income are dropped. While remittance data in surveys can be subject to measurement error, the use of survey fixed effects will capture any common survey-level effects, and there is no strong reason to believe such measurement error would be correlated with education status. Mean and median reported remittances also seem to be of the right order of magnitude when compared with other surveys and migrant incomes. The sample weights provided with the data are always used. Data are pooled by poststratifying by country of birth and by education, so that the combined weighted observations match the distribution of developing country migrants to all OECD countries in 2000 (Docquier and Marfouk2005). The supplemental appendix provides further details. Table 2 presents summary statistics for each country survey and the pooled samples of all destination countries. Overall, 37 percent of migrants in the database have completed a university degree, ranging from 4 percent in the Spanish Netherlands Interdisciplinary Demographic Institute (NIDI) survey to 6. Exceptions include longitudinal surveys of immigrants from Canada and New Zealand, which can only be accessed through datalabs in these countries, and so are not included here. 7. High income countries are defined based on the World Bank Country Classification Code, April 2009.

TABLE 1. Migrant Datasets Dataset Name Year Number of observations a Population Methodology Australia LSIA Belgium IRSHS France 2MO France DREES Germany SOEP Italy NIDI Longitudinal Survey of Immigrants to Australia International Remittance Senders Household Survey Survey of Households Transfer of Funds to their Countries of Origin Profile and Tracking of Migrants Survey German Socio-Economic Panel Study Netherlands Interdisciplinary Demographic Institute International Migration Survey Survey of Brazilians and Peruvians in Japan Consumentenbond Survey of Remittances 1997 2,537 Primary applicant migrant arrivals September 1993 August 1995 2005 377 Immigrants from DR Congo, Nigeria, and Senegal 2007 713 Remitters to Algeria, Morocco, Tunisia, Turkey and the countries of Sub-Saharan Africa 2006 4,278 New and regularized migrants with 1þ year residence permits 2000 854 Resident population of the Federal Republic of Germany in 1984. 1997 1,072 Egyptians and Ghanaians who immigrated within past 10 years Japan IDB 2005 846 Latin American immigrant adults living in Japan Netherlands CSR 2005 648 Major immigrant populations: Moroccans, Turks, Surinamese, Antilleans, Somalis, and Ghanaians Norway LKI Living Conditions of Immigrants 1996 2,466 Immigrants from 10 countries: Survey Bosnia and Herzegovina, Chile, Iraq, Iran, Pakistan, Serbia, Somalia, Sri Lanka, Turkey, and Vietnam Spain ENI National Survey of Immigrants 2006 9,234 Foreign-born who (intend to) live in Spain for 1þ years Sample of official records of those living in cities Referrals through Embassies. Interviews of remitters at post offices in high-migrant regions Sample of official records Sample of official records Interviews at migrant meeting places Interviews in 15 cities Face-to-face interviews Representative survey of immigrant population from these countries Sample of official neighborhood rosters (Continued) Bollard, McKenzie, Morten, and Rapoport 139

TABLE 1. Continued Dataset Name Year Spain NIDI Netherlands Interdisciplinary Demographic Institute International Migration Survey Black/Minority Ethnic Remittance Survey Number of observations a Population Methodology 1997 1,020 Moroccans and Senegalese who immigrated within past 10 years UK BME 2006 993 Migrant minorities who have remitted in past 12 months U.S. NIS New Immigrant Survey 2003 7,046 Migrants receiving green cards May November 1993 Pew National Survey of Latinos 2006 1084 Nationally representative sample of Latino respondents ages 18 and older U.S. Pew Pew National Survey of Latinos 2006 1,084 Nationally representative sample of Latino respondents ages 18 and older Geographical sampling, & references from sampled Sampling of geographical blocks Sample of official records Sampled phone numbers in high-latino areas Sampled phone numbers in high-latino areas 140 THE WORLD BANK ECONOMIC REVIEW Note: Number of observations used to calculate the first result in each column of table 2.

TABLE 2. Survey Means, by Education Australia Belgium France France Germany Italy Japan Netherlands Norway Spain Spain UK U.S. U.S. Pooled Pooled Pooled Variable and education level LSIA IRSHS 2MO DREES SOEP NIDI IDB CSR LBI ENI NIDI BME NIS Pew Extensive Intensive Total Number of observations 2,656 451 717 4,280 900 1,153 1,065 836 2,466 10,282 1,113 1,152 7,352 1,304 33,022 28,981 26,276 Fraction with university education 0.32 0.59 0.07 0.18 0.20 0.21 0.14 0.11 0.12 0.23 0.04 0.20 0.34 0.06 0.36 0.37 0.37 Total remittances ($ per year) No university 286 1,681 1,380 368 2,724 2,662 993 988 3,099 2,691 375 1,530 699 793 650 University 379 2,475* 1,652 511 2,227 2,920 1,405* 743** 2,835 2,629 1,145** 671** 868 897 874 Fraction who remit No university 0.41 0.91 0.23 0.18 0.60 0.80 0.34 0.49 0.78 0.15 0.54 0.31 0.32 0.32 University 0.37 0.86 0.23 0.20 0.45** 0.90** 0.29 0.37** 0.48** 0.17 0.43 0.27** 0.27** 0.27** Log remittances No university 5.78 6.92 6.62 6.97 7.89 7.76 6.49 7.15 7.99 6.77 7.01 7.34 6.96 6.82 6.91 University 6.23** 7.29** 6.92 7.01 8.11 7.70 6.81** 7.22 8.49* 6.92 7.40** 6.97 7.02 6.97* 7.00 Household income ($ per year) No university 14,457 16,918 23,173 18,612 19,526 10,903 34,014 32,467 14,066 9,074 44,631 33,297 22,417 22,624 23,583 21,964 University 13,556 25,534** 31,301* 28,674** 21,984 13,302* 43,624** 41,995** 19,914** 10,168 50,565 61,084 34,729** 38,948** 38,669** 39,087** Log income No university 9.5 9.5 9.8 9.6 9.8 9.3 10.2 10.1 9.4 9.0 10.3 9.2 9.7 9.6 9.5 9.5 University 9.8** 9.8** 10.0 9.9** 9.8 9.4 10.4 10.3** 9.7** 9.2 10.4 10.0** 10.2** 9.9** 9.9** 9.9** Working No university 0.48 0.70 0.87 0.80 0.63 0.82 0.93 0.48 0.68 0.81 0.82 0.66 0.66 0.65 0.66 0.64 University 0.67** 0.74 0.86 0.86** 0.67 0.87 0.93 0.70** 0.73** 0.66 0.90** 0.78** 0.77* 0.75** 0.74** 0.73** Household size No university 3.81 1.88 2.51 2.90 1.80 1.53 3.82 1.84 3.33 4.10 3.44 3.73 3.76 University 3.44** 2.55** 1.90** 2.58 2.16** 1.76** 3.19** 1.95 3.04* 3.49** 3.17** 3.35** 3.36** Married No university 0.73 0.72 0.65 0.67 0.61 0.56 0.47 0.64 0.66 0.54 0.63 0.63 0.63 University 0.80** 0.51** 0.71* 0.59 0.60 0.48* 0.56** 0.51 0.86** 0.56 0.73** 0.74** 0.74** Spouse outside country No university 0.03 0.25 0.05 0.06 0.42 0.05 0.05 0.05 0.06 University 0.01* 0.19 0.01** 0.05 0.10** 0.03** 0.03** 0.03** 0.03** Number of children No university 1.29 1.16 1.78 1.06 2.50 2.06 1.58 2.25 2.37 1.99 2.05 2.03 University 1.22 0.89** 1.27** 1.00 2.15** 1.85** 0.62** 1.35** 1.81** 1.37** 1.37** 1.37** Children outside country No university 0.21 0.10 0.25 0.71 0.20 0.16 0.38 1.10 0.73 0.49 0.45 0.48 0.50 University 0.07** 0.06 0.17** 0.49* 0.15 0.09 0.26** 0.21** 0.31** 0.37 0.24** 0.25** 0.25** Number of parents No university 1.97 1.13 0.95 1.35 1.42 1.27 2.18 1.81 1.84 1.83 (Continued) Bollard, McKenzie, Morten, and Rapoport 141

TABLE 2. Continued Australia Belgium France France Germany Italy Japan Netherlands Norway Spain Spain UK U.S. U.S. Pooled Pooled Pooled Variable and education level LSIA IRSHS 2MO DREES SOEP NIDI IDB CSR LBI ENI NIDI BME NIS Pew Extensive Intensive Total University 2.32** 1.03 0.70** 1.32 1.35** 1.37 2.74** 2.18** 2.21** 2.23** Parents outside country No university 1.48 0.81 0.42 0.94 1.03 1.01 1.23 0.88 0.98 0.98 1.00 University 2.00** 0.88 0.54 0.67** 1.17* 1.04 1.33 1.26** 1.30** 1.31** 1.31** Years spent abroad No university 3.70 9.32 17.90 4.00 19.20 6.69 8.35 18.46 10.06 7.27 14.89 7.35 16.43 9.20 11.17 10.29 University 3.91** 12.28** 12.70** 4.21 13.51** 7.02 9.18 19.36 12.41** 6.74 14.66 7.05 18.34 8.06** 8.75** 8.40** Legal immigrant No university 1.00 1.00 0.84 0.51 0.66 1.00 0.87 0.84 0.85 University 1.00 1.00 0.85 0.39** 0.82* 1.00 0.85** 0.84 0.84 Will return home No university 0.02 0.45 0.06 0.23 0.39 0.01 0.08 0.35 0.63 0.09 0.19 0.09 0.16 0.11 University 0.04 0.65** 0.10* 0.17 0.53** 0.02 0.08 0.51 0.70 0.13** 0.14 0.09 0.12** 0.09* *Significant at the 5 percent level ** significant at the 1 percent level Note: Significant results indicate that the mean of the variable is statistically different between university-educated and non-university educated households. See table 1 for full survey names. Source: Authors analysis based on data described in the text. 142 THE WORLD BANK ECONOMIC REVIEW

Bollard, McKenzie, Morten, and Rapoport 143 59 percent in the Belgium IRSHS. The table also summarizes the covariates by the maximum educational attainment of all adult migrants in the household. Altogether, including both the extensive and intensive margins, more highly educated migrants send home an average of $874 annually, compared with $650 for less educated migrants. There are two opposing effects of education: negative on the extensive margin, and positive on the intensive margin. At the extensive margin, migrants with a university degree are less likely to remit anything than migrants without a degree: 32 percent of low-skilled migrants send some money home, compared with 27 percent of university-educated migrants. However, conditional on remitting (the intensive margin), highly educated migrants send about 9 percent more than do less educated migrants. Characteristics that can affect remittance behavior differ between less and more educated migrants. First, more skilled migrants are both more likely to live in a household with working adults and to have a higher household income than are low skilled migrants. But contrary to conventional wisdom, household composition does not differ much for migrants by education level: on average, only 6 percent of low skilled migrants have a spouse outside the country, compared with 3 percent of high skilled migrants. Low skilled migrants are significantly less likely to be married (63 percent) than are high skilled migrants (74 percent). Low skilled migrants have more children (an average of 2.03 compared with 1.37 for high skilled migrants), as well as more children living outside the destination country (0.50) than do high skilled migrants (0.25). However, low skilled migrants also have more family inside the destination country than do high skilled migrants: the average household size for low skilled migrants is 3.76 people, statistically different from the mean household size of 3.36 people for high skilled migrants. 8 Another piece of conventional wisdom, that more educated people are less likely to return home, is also not supported by the microdata. Indeed, more educated migrants have spent less time abroad (mean of 8.4 years) than have less educated migrants (10.3 years). Reported plans to return home are similar between the two groups: 9 percent of high skilled migrants report planning to return home, compared with 11 percent of low skilled migrants. While one should be cautious with treating both measures as truly reflecting return probabilities; at the very least, they do not indicate a strong tendency for the low skilled to be more likely to return. The simple comparison of means in table 2 shows differences in remittance behavior by education status. However, these comparisons show only that more educated developing country emigrants remit more than less educated developing country emigrants. This risks confounding differences in remittance 8. In some cases this might reflect households in which poorer, less skilled migrants live with other immigrants who are not family members. The database can identify the presence of a spouse, child, or parent in the home country household but cannot identify who migrants live with abroad or the extent to which they share resources within the household abroad.

144 THE WORLD BANK ECONOMIC REVIEW behavior among migrants from different countries with differences in remittance behavior by education level: the next section aims to separate these two differences. III. RESULTS:THE R ELATIONSHIP BETWEEN E DUCATION AND R EMITTANCES Results are reported for regressions of three remittance measures on education: total remittances (both extensive and intensive margins), an indicator for having remitted in the previous year (extensive margin), and log total remittances conditional on remitting (intensive margin; table 3). All regressions include country of birth fixed effects and dataset fixed effects. The key result is that more educated migrants remit more. In the pooled sample, migrants with a university degree remit $298 more per year than nonuniversity educated migrants (row last, last column), with a mean annual remittance for all migrants of $734. This overall effect is composed of a negative (statistically insignificant) effect at the extensive margin and a highly significant positive effect at the intensive margin. The results are consistent when the second measure of education, years of schooling, is considered. Results for individual countries are mixed at the extensive margin, with education significantly positively associated with the likelihood of remitting in two surveys (the U.S. NIS and the Survey of Brazilians and Peruvians in Japan), significantly negatively associated with this likelihood in three surveys (the U.S. Pew survey and both Spanish surveys), and no significant relationship in the other six surveys, with three positive and three negative point estimates. One general observation is that a more negative relationship appears in surveys that focus on sampling migrants through community-sampling methods, such as the NIDI surveys, which take their sample from places where migrants cluster, and the Pew Hispanic surveys, which randomly dial phone numbers in areas with dense Hispanic populations. One might expect that educated migrants who live in such areas (and who take the time to respond to phone or on-the-street surveys) would be less successful than educated migrants who live in more integrated neighborhoods and thus who would not be picked up in these surveys. In contrast, at the intensive margin, 10 of 12 individual surveys show a positive relationship between remittances and education, 5 of them statistically significant, and 2 show a negative and insignificant relationship. Thus it is not surprising that when the data are pooled there is a strong positive association at the intensive margin and that it outweighs the small negative and insignificant relationship at the extensive margin in the total effect. This point is made graphically on a log scale in figure 1, which plots the nonparametric relationship between total remittances and years of schooling, after linearly controlling for dataset fixed effects using a partial linear model (Robinson 1988), together with a 95 percent confidence interval. The vertical lines demarcate the quartiles of years of schooling. Average remittances steadily

TABLE 3: Coefficients from Regressions of Remittance Measures on Education Australia Belgium France France Germany Italy Japan Netherlands Norway Spain Spain UK U.S. U.S. Pooled Pooled Pooled Variable LSIA IRSHS 2MO DREES SOEP NIDI IDB CSR LBI ENI NIDI BME NIS Pew Extensive Intensive Total Education measured by university degree Total remittances ($ per year) 58.4 922.8** 291.0-526.6 237.5-92.6-168.8 769.5** -554.0* 298.0* Number of observations 2,537 377 854 1,072 846 9,234 1,020 7,046 1,084 24,033 Extensive: Remits indicator 20.019-0.055 0.014 0.042-0.065 0.091** 0.012-0.049** -0.232** 0.038** -0.140* -0.018-0.010 Number of observations 2,654 451 4,278 854 1,153 1,030 2,466 10,282 1,112 7,113 1,296 32,651 25,907 Intensive: Log remittances 0.341* 0.433** 0.363 0.492 0.073-0.057 0.333** 0.093 0.430* 0.168 0.397* -0.199 0.249** 0.226** Number of observations 958 317 713 184 545 690 648 3,966 761 993 1,118 514 11,392 9,038 Education measured in years Total remittances ($ per year) 19.08* 86.50 26.39-7.56-3.03 2.40-13.65 86.53 64.89 57.81 Number of observations 2,531 377 854 1,072 846 9,164 1,020 7,033 1,084 23,944 Extensive: Remits indicator 0.0080-0.0042 0.0018 0.0145 0.0010 0.0024** 0.0008-0.0023-0.0072** 0.0034** 0.0010 0.0006 0.0014 Number of observations 2,648 451 5,529 854 1,153 1,030 2,450 10,201 1,112 7,100 1,296 32,535 25,807 Intensive: Log remittances 0.0441* 0.0341 0.0224* -0.0085-0.0032-0.0040 0.0247* 0.0199** 0.0091 0.0548* 0.0329* 0.0369 0.0256** 0.0229** Number of observations 956 317 713 184 545 690 648 3,942 761 993 1,116 514 11,364 9,010 Means Total remittances ($ per year) 316 2,159 1,399 396 2,621 2,692 1,040 932 3,089 2,679 633 1,479 764 2,466 734 Fraction who remit 0.40 0.85 0.23 0.19 0.53 0.77 0.34 0.41 0.75 0.15 0.46 0.30 1.00 0.27 Fraction with university 0.32 0.60 0.07 0.18 0.20 0.21 0.12 0.11 0.12 0.23 0.04 0.20 0.33 0.06 0.36 0.31 0.38 Years of education 13.4 14.2 7.7 12.0 11.5 14.1 13.3 10.7 12.2 11.4 7.5 13.4 13.4 9.4 12.9 12.3 13.0 *Significant at the 5 percent level ** significant at the 1 percent level Note: See table 1 for full survey names. Source: Authors analysis based on data described in the text. Bollard, McKenzie, Morten, and Rapoport 145

146 THE WORLD BANK ECONOMIC REVIEW Figure 1. Total Remittances by Years of Schooling Note: Figure depicts a semiparametric regression line from a partial linear model with dataset dummy variable evaluated at means; 95 percent pointwise confidence intervals shown from 500 bootstrap repetitions. Vertical lines separate quartiles. Source: Authors analysis based on data described in the text. increase from around $500 in the lowest education quartile to close to $1,000 for those with university degrees. Moreover, the positive association increases most strongly for migrants with postgraduate education, which shows that not only do migrants with some university education remit more than those without, but also that migrants with postgraduate degrees remit more than those with only a couple of years of university. Robustness Although this database on remittances is the most comprehensive available, there are clear limitations, which make it important to see how sensitive the results are to alternative ways of using these surveys. First note that the results pertain only to migration in a sample of OECD countries. The surveys cover a large share of OECD destinations, but they omit other important destinations for developing country migrants such as the Gulf countries and South Africa. This is a limitation shared by the macro studies (Faini 2007; Niimi and others 2008), which also have data only for migrants in OECD countries. Nevertheless, the same forces acting on migrants in the OECD countries are likely to apply in these other destinations: more educated migrants will earn higher incomes and therefore remit more. Although data are rare, there is some evidence to support this is in a study of Pakistani migrants in the Gulf countries, which found that conditional on age and duration of

Bollard, McKenzie, Morten, and Rapoport 147 stay, more educated Pakistani migrants remitted more (Abbasi and Hashmi 2000). Moreover, it is still the case that there are a large number of low skilled migrants in the OECD. A large majority of migrants in the pooled sample (63 percent) do not have a university education. A reasonable concern is whether surveys like the U.S. NIS, which capture only legal immigrants, are missing most of the low-skilled migrants. Comparing the skill distribution of immigrants included in the NIS with that of immigrants included in the U.S. Census (which is generally believed to do a good job surveying both legal and illegal migrants), does show a higher skill level in the NIS (12.26 mean years of education) than in the Census (10.84 years). However, once Mexican immigrants are excluded (the group with the largest number of illegal immigrants), the skill distributions of the NIS (12.96 mean years of education) and the Census (12.21) are much closer, and16 percent of immigrants in both the NIS and the Census have 8 years of education or less. The first two columns of table 4 then show that the results for the association between remittances and education continue to hold in the NIS (and, if anything, are more strongly positive) when Mexican immigrants are excluded (table 4, columns 1 and 2). Columns 3 and 4 show that this is also true for the pooled sample of all surveys, which suggests that failure to capture illegal migrants in the survey is not driving the main result. A second potential concern is whether it is valid to pool so many different surveys with different sampling methods and differing degrees of representativeness. Note that survey fixed effects are included in the regression analysis, so that only within-survey variation is used to identify the effect of education; the pooled estimate is thus a consistent estimate for the average association among the surveys. Nonetheless, as an alternative, the regressions are run only for the five surveys based on representative sampling from a list of migrants: (Longitudinal Survey of Immigrants to Australia (LSIA), the French Profile and Tracking of Migrants Survey (DREES), the German Socioeconomic Panel Study (SOEP), and the Spanish ENI and the NIS). The results show point estimates and levels of statistical significance that are very close to those for the full pooled sample (see table 4, column 5). This demonstrates that the results are not being driven by the specialized surveys of particular migrant groups, such as the Japanese and Belgian surveys. Finally, one might query whether the results are being driven by students. That could influence the results based on university education if there were many students studying for undergraduate degrees who do not send remittances and do not yet have a college degree. There are several reasons to believe that this is not the main factor driving results. First, the LSIA and NIS surveys do not include students, which eliminates from the sample students in the countries that are among the most popular destinations for international study. Second, many international students come for postgraduate education, so they would be classified as having a college education and remitting little, which

TABLE 4. Robustness Checks Variable U.S. New Immigrant Survey(NIS) sample Full sample Excluding Mexicans Full sample Excluding NIS Mexicans Pooled sample Nationally representative samples Only migrants ages 25 þ Only working migrants Education measured by university degree Total remittances ($ per year) 769.5** 839.7** 298.0* 306.8* 318.5* 267.7 308.7 Number of observations 7,046 5,922 24,033 22,909 19,643 21,343 16,693 Extensive: Remits indicator 0.038** 0.046** -0.010 0.000-0.011-0.023* 0.048** Number of observations 7,113 5,984 25,907 24,778 20,875 23,043 18,147 Intensive: Log remittances 0.397* 0.458** 0.226** 0.236** 0.220** 0.192** 0.230** Number of observations 1,118 982 9,038 8,902 6,220 8,303 7,360 Education measured in years Total remittances ($ per year) 86.53 108.09 57.81 68.17 62.52 58.78 78.49 Number of observations 7,033 5,909 23,944 22,820 19,554 21,263 16,639 Extensive: Remits indicator 0.0034** 0.0021 0.0014 0.0016 0.0017 0.0007 0.0031* Number of observations 7,100 5,971 25,807 24,678 20,775 22,954 18,083 Intensive: Log remittances 0.0329* 0.0391* 0.0229** 0.0260** 0.0307** 0.0235** 0.0242** Number of observations 1,116 980 9,010 8,874 6,192 8,279 7,335 Means Total remittances ($ per year) 633 719 734 813 614 772 935 Fraction who remit 0.15 0.16 0.27 0.29 0.24 0.28 0.33 Fraction with university 0.33 0.40 0.38 0.38 0.38 0.38 0.38 Years of education 13.4 14.0 13.0 13.1 13.1 13.0 13.4 148 THE WORLD BANK ECONOMIC REVIEW *Significant at the 5 percent level ** significant at the 1 percent level. Note: Nationally representative surveys are Australia LSIA, France DREES, Germany SOEP, Spain ENI and U.S. NIS. See table 1 for full survey names. Source: Authors analysis based on data described in the text.

Bollard, McKenzie, Morten, and Rapoport 149 would offset any effect of undergraduates. 9 As a final check, the analysis is restricted to individuals who are working (table 4, last column). Since more educated individuals are more likely to be working, this eliminates one channel through which the more educated can earn more and thereby remit more. Nevertheless, even with this restriction, there is a significant positive coefficient at both the extensive and intensive margins, and the point estimate for total remittances is similar in magnitude, although it is not statistically significant. Taken together, these results indicate that the basic finding of a positive relationship between total remittances and education appears reasonably robust to alternative ways of combining the surveys. Channels This section uses these microdata to explore some of the channels through which education might influence remittances. Proxies are added to the model to control for differences in household income and work status, in household demographics and the presence of family abroad, in time spent abroad, in legal status, and in intentions to return home. Table 5 shows the results of adding this full set of variables to the pooled model, using years of education as the measure of educational attainment. These channels operate as theory would predict. Households with more income and with adults who work more are more likely to remit: households where a migrant member is working send $345 more annually, with an extra $38 remitted annually for each 10 percent increase in income. As expected, family composition variables are also strongly significant both overall and for the extensive and intensive margins: a spouse outside the country is associated with a colossal additional $1,120 remitted each year, approximately one and a half times the mean annual remittance for all migrants. Each child living outside the destination country is associated with an additional $340 remitted annually and each parent for an additional $180. Residing in the destination country legally is associated with an additional $400 annually, providing no evidence that legal migrants lose their desire to remain in contact with their home country. Migrants who plan to move back home also remit significantly more, but this effect is primarily through the extensive margin rather than the intensive margin. Which channels account for the association between education and remittance behavior? Tables 6, 7, and 8 report how the coefficient on education in an ordinary least squares regression changes as controls are added for total remittances, the extensive margin, and the intensive margin. Each panel of each table first shows the baseline education coefficient from regressing remittances only on education and country of birth and dataset fixed effects (from table 3). Each succeeding row then shows changes in this coefficient when controls are 9. In the United States, 47 percent of international students are studying for postgraduate degrees, compared with 12 percent for associate degrees and 32 percent for bachelor degrees (http://opendoors.iienetwork.org/?p=150827).

150 THE WORLD BANK ECONOMIC REVIEW TABLE 5. Remittances on Years of Education for Pooled Sample with All Controls Total Extensive Intensive Variable remittances Remits Log remittances Years of education 37.81 20.002* 0.017** (29.64) (0.001) (0.005) Log income 384.59** 0.023** 0.364** (105.37) (0.003) (0.034) Working 345.06** 0.113** 0.514** (90.80) (0.010) (0.065) Household size 28.14 20.002 0.015 (17.67) (0.002) (0.016) Married 289.77 0.004 20.097 (68.78) (0.010) (0.061) Spouse outside country 1,120.95** 0.145** 0.568** (236.04) (0.020) (0.097) Number of children 2121.56** 20.006 20.099** (36.44) (0.003) (0.027) Children outside country 337.78** 0.048** 0.228** (75.14) (0.006) (0.039) Number of parents 247.07 20.020** 20.125** (53.56) (0.005) (0.045) Parents outside country 182.58** 0.063** 0.243** (38.02) (0.006) (0.045) Years spent abroad / 100 2,539.77 0.251** 1.744** (2,533.08) (0.095) (0.656) Years spent abroad squared / 100 231.43 20.010** 20.033* (27.14) (0.002) (0.015) Legal immigrant 398.79** 0.096** 0.167** (121.36) (0.018) (0.061) Will return home 692.30** 0.095** 0.085 (201.83) (0.021) (0.072) Number of observations 23,944 32,535 11,364 *Significant at the 5 percent level ** significant at the 1 percent level. Note: Includes dummy variables for missing covariates and fixed effects for country of birth and survey. Trimmed remittances greater than twice income. Pooled samples poststratified by country and education. Source: Authors analysis based on data described in the text. added for income and work status, family composition, and all controls from table 5 (income and family controls, as well as legal status, time spent abroad, and intent to return home). Remittance behavior is accounted for primarily by income and not by differences in family composition. The baseline result for total remittances from table 3, controlling only for country of birth and dataset fixed effects, is that migrants with a university degree remit $300 more than migrants without one. Controlling for the full set of covariates (the all row) reduces the coefficient on university degree by two-thirds, and it becomes statistically insignificant. The third row adds just the family composition variables to the baseline

TABLE 6. Education Coefficient as Controls Are Added: Total Annual Remittances (U.S. dollars) Australia Belgium Germany Italy Japan Spain Spain U.S. U.S. Pooled Variable LSIA IRSHS SOEP NIDI IDB ENI NIDI NIS Pew total University education Baseline 58.4 922.8** 291.0 2526.6 237.5 292.6 2168.8 769.5** 2554.0* 298.0* (61.1) (351.4) (275.6) (411.6) (374.1) (62.8) (749.4) (254.4) (227.2) (137.6) Income 210.1 557.0* 238.5 2623.9 166.5 2189.3** 24.7 396.6* 2741.5** 102.3 (62.4) (281.4) (262.2) (407.2) (359.8) (63.5) (729.0) (174.4) (263.8) (92.8) Family a 29.8 534.7 237.8 2306.7 317.5 2112.8 26.9 623.6** 2698.6** 228.2* (61.4) (310.5) (243.5) (394.7) (380.3) (57.6) (725.9) (204.7) (241.9) (103.1) All 216.5 475.8 144.6 2539.6 328.7 2181.7** 266.2 402.2** 2835.7** 99.9 (62.1) (272.7) (179.8) (383.3) (365.3) (58.6) (698.6) (154.3) (269.9) (71.6) Number of observations 2,537 377 854 1,072 846 9,234 1,020 7,046 1,084 24,033 Years of education Baseline 19.08* 86.50 26.39 27.56 23.03 2.40 213.65 86.53 64.89 57.81 (9.01) (45.11) (29.37) (34.05) (7.92) (7.36) (19.95) (46.50) (44.97) (37.08) Income 7.99 47.80 3.51 232.44 22.59 213.39 226.95 44.98 49.18 32.12 (8.69) (38.28) (27.33) (33.39) (11.50) (7.41) (19.68) (40.00) (45.09) (31.98) Family a 17.03 29.28 25.56 47.31 21.86 3.93 10.32 80.78 47.95 55.43 (8.98) (38.45) (27.79) (34.93) (8.62) (6.84) (19.98) (44.75) (46.37) (34.24) All 8.86 33.77 9.66 22.64 1.99 27.57 4.50 54.81 27.01 37.81 (8.91) (36.94) (22.82) (32.79) (10.63) (6.84) (19.32) (37.32) (46.38) (29.64) Number of observations 2,531 377 854 1,072 846 9,164 1,020 7,033 1,084 23,944 *Significant at the 5 percent level ** significant at the 1 percent level. Note: Baseline row includes only country of birth and dataset fixed effects. Income row adds working dummy and log income to baseline. Family row adds seven family member controls to baseline. All row is full specification from table 3. Trimmed remittances greater than twice income. Pooled samples poststratified by country and education. See table 1 for full survey names. a. Includes household size, dummy variable if married, dummy variable if spouse is outside the country, number of children, number of children outside the country, number of parents, and number of parents outside the country Source: Authors analysis based on data described in the text. Bollard, McKenzie, Morten, and Rapoport 151